Tag: Artificial Intelligence

Embedding AI Error Detection Into Primary Care Safety Culture

As artificial intelligence (AI) become increasingly embedded in routine healthcare – supporting tasks such as triage, documentation, interpretation of investigation, diagnosis and patient communication – it introduces new patient safety risks through incorrect outputs (“hallucinations”) that should be treated as safety errors rather than technical glitches. In our article in the Journal of Patient Safety, we argue that primary care must extend its established safety culture to AI by systematically detecting, classifying, reporting, and learning from AI-related errors using principles already applied to human error, such as audit, governance, and incident reporting.

We highlight evidence that AI-generated clinical text can contain omissions, fabrications, or unsafe recommendations that may not be apparent to clinicians and patients and that risk becoming “silent errors” in electronic health records. These errors can then contribute to cognitive offloading if clinicians over-trust AI outputs. To mitigate these risks, we call for routine AI oversight in practice (including review, sampling, and escalation), explicit clinician accountability for AI-influenced outputs, patient engagement in spotting discrepancies, and closer collaboration with AI developers.

Ultimately, AI errors are inevitable, and that embedding AI safety as a core, proactive design feature – rather than an afterthought – is essential to ensure AI enhances rather than compromises patient safety in primary care.

Relevance Over Recall: Rethinking How AI Uses Clinical Data

Our article in the Journal of the Royal Society of Medicine argues that safe and effective AI in healthcare must incorporate mechanisms that emulate human judgement – down-weighting old, inaccurate or superseded information and prioritising what is recent, clinically relevant and reaffirmed – so that AI supports, rather than disrupts, high-quality patient care.

Clinicians constantly revise, reinterpret and filter past information so that only what is relevant, accurate and timely shapes present-day management decisions; medical records function as dynamic “working tools” rather than fixed archives. By contrast, many AI systems lack this capacity for selective forgetting and often treat all historical data as equally meaningful.

This can lead to outdated or low-confidence diagnoses being repeatedly resurfaced, persistent labels influencing clinical expectations, and irrelevant, long-resolved events cluttering summaries and decision-support outputs. Such indiscriminate recall not only risks misdirecting clinical care, but also adds to information overload, exacerbates cognitive burden and contributes to clinician burnout. Importantly, it can also undermine patient trust when obsolete or stigmatising terms continue to shape interactions with the clinicians and the healthcare system.

Bridging the Equity Gap in AI Healthcare Diagnostics

In an era where artificial intelligence (AI) is rapidly reshaping the landscape of healthcare diagnostics, our recent BMJ article sheds light on a critical issue: the equity gap in AI healthcare diagnostics. The UK’s substantial investment in AI technologies underscores the nation’s commitment to enhancing healthcare delivery through innovations. However, this evolution brings to the forefront the need for equity: defined as fair access to medical technologies and unbiased treatment outcomes for all.

AI’s potential in diagnosing clinical conditions like cancer, diabetes, and Alzheimer’s Disease is promising. Yet, the challenges of data representation, algorithmic bias, and accessibility of AI-driven technologies loom large, threatening to perpetuate existing healthcare disparities. Our article highlights that the quality and inclusivity of data used to train AI tools are often problematic, leading to less representative data and biases in AI models. These biases can adversely affect diagnostic accuracy and treatment outcomes, particularly for people from ethnic minority groups and women, who are often under-represented in medical research.

To bridge this equity gap, we advocate for a multi-dimensional systems approach rooted in strong ethical foundations, as outlined by the World Health Organization. This includes ensuring diversity in data collection, adopting unbiased algorithms, and continually monitoring and adjusting AI tools post-deployment. We also suggest establishing digital healthcare testbeds for systematic evaluation of AI algorithms and promoting community engagement through participatory design to tailor AI tools to diverse health needs.

A notable innovation would be the creation of a Health Equity Advisory and Algorithmic Stewardship Committee, spearheaded by national health authorities. This committee would set and oversee compliance with ethical and equity guidelines, ensuring AI tools are developed and implemented conscientiously to manage bias and promote transparency.

The advancement of AI in healthcare diagnostics holds immense potential for improving patient outcomes and healthcare delivery. However, realising this potential requires a concerted effort to address and mitigate biases, ensuring that AI tools are equitable and representative of the diverse populations they serve. As we move forward, prioritising rigorous data assessment, active community engagement, and robust regulatory oversight will be key to reducing health inequalities and fostering a more equitable healthcare landscape through the use of AI in healthcare diagnostics.

Dr Demis Hassabis, Co-Founder and CEO of DeepMind, Speaks about AI in Healthcare

OnOn 28 September 2017, I attended the Annual Institute of Global Health Innovation Lecture: Artificial General Intelligence and Healthcare, delivered by Dr Demis Hassabis, co-founder and CEO of Google DeepMind. Artificial intelligence is the science of making machines smart argued Dr Hassabis, so how can we make it improve the healthcare sector? Dr Hassabis then went on to describe the work that DeepMind was carrying out in healthcare in areas such as organising information, deep learning to support the reporting of medical images (such as scans and pathology slides), and biomedical science. Dr Hassabis also discussed the challenges of applying techniques such as reinforcement learning in healthcare. He concluded that artificial intelligence has great scope for improving healthcare; for example, by prioritising the tasks that clinicians had to carry out and by providing decision support aids for both patients and doctors.